An internetwork is a collection of computer networks interconnected by nodes, each such node may be a general-purpose computer or a specialized device, such as a router. As such, an internetwork is often called a network of networks. The purpose of building an internetwork is to provide information services to end nodes, each end node may be a general-purpose computer or a specialized device, such as a camera or a display. The Internet is an internetwork in which information is organized into packets to be distributed on a store- and forward manner from source to destination end nodes, and in which routers and end nodes use the Internet Protocol (IP) to communicate such packets.
The World Wide Web (also known as WWW or Web) has become an essential information service in the Internet. The Web constitutes a system for accessing linked information objects stored in end nodes (host computers) all over the Internet. Berners-Lee wrote the original proposal for a Web of linked information objects (T. Berners-Lee, “Information Management: A Proposal,” CERN Document, March 1989). The Web consists of a vast collection of information objects organized as pages, and each page may contain links to other pages or, more generally, information objects with which content is rendered as audio, video, images, text or data. Pages are viewed by an end user with a program called a browser (e.g., Netscape Navigator™). The Web browser runs in an end system at the user premises. The client (Web browser) obtains the required information objects from a server (Web server) using a request-response dialogue as part of the Hypertext Transfer Protocol (HTTP). Information objects are identified by means of names that are unique throughout the Internet; these names are called Uniform Resource Locators or URLs. A URL consists of three components:                (1) the protocol or scheme to be used for accessing the object (e.g., http);        (2) the name (a DNS name) of the host on which the object is located; and        (3) a local identifier that is unique in the specified host.        
Like any large-scale system, the Web requires the use of mechanisms for scaling and reliability. More specifically, as the number of information objects that can be obtained through the Web increases, people find it more difficult to locate the specific information objects they need. Furthermore, as the number of Web users and servers increase, the sites or servers that store the requested information objects may be very far from the users requesting the objects, which leads to long latencies in the access and delivery of information, or the servers storing the information objects may be overwhelmed with the number of requests for popular information objects.
To enable the Web to scale to support large and rapidly increasing numbers of users and a vast and growing collection of information objects, the information objects in the Web must be stored distributedly at multiple servers, in a way that users can retrieve the information objects they need quickly and without overwhelming any one of the servers storing the objects. Accordingly, distributing information objects among multiple sites is necessary for the Web to scale and be reliable. The schemes used to accomplish this are called Web caching schemes. In a Web caching scheme, one or multiple Web caches or proxy Web servers are used in computer networks and the Internet to permit multiple host computers (clients) to access a set of information objects from sites other than the sites from which the content (objects) are provided originally. Web caching schemes support discovering the sites where information objects are stored, distributing information objects among the Web caches, and retrieving information objects from a given Web cache. The many proposals and implementations to date differ on the specific mechanisms used to support each of these services.
Many methods exist in the prior art for determining the server, cache, minor server, or proxy from which information objects should be retrieved. The prior art dates to the development of the ARPANET in the 1970s and the study and implementation of methods to solve the file allocation problem (FAP) for databases distributed over the ARPANET and computer networks in general.
File allocation methods for distributed databases (e.g., W. W. Chu, “Optimal File Allocation in a Multiple Computer System,” IEEE Transactions on Computers, October 1969; S. Mahmoud and J. S. Riordon, “Optimal Allocation of Resources in Distributed Information Networks,” ACM Transactions on Data Base Systems, Vol. 1, No. 1, March 1976; H. L. Morgan and K. D. Levin, “Optimal Program and Data Locations in Computer Networks,” Communications of the ACM, Vol. 20, No. 5, May 1977) and directory systems (e.g., W. W. Chu, “Performance of File Directory Systems for Data Bases in Star and Distributed Networks,” Proc. National Computer Conference, 1976, pp. 577-587; D. Small and W. W. Chu, “A Distributed Data Base Architecture for Data Processing in a Dynamic Environment,” Proc. COMPCON 79 Spring) constitute some of the earliest embodiments of methods used to select a delivery site for accessing a file or information object that can be replicated at a number of sites.
Another example of this prior art is the method described by Chiu, Raghavendra and Ng (G. Chiu, C. S. Rahgavendra, and S. M. Ng, “Resource Allocation with Load Balancing Consideration in Distributed Computing Systems,” Proc. IEEE INFOCOM 89, Ottawa, Ontario, Canada, April 1989, pp. 758-765). According to this method, several identical copies of the same resource (e.g., a file, an information object) are allocated over a number of processing sites (e.g., a mirror server, a cache) of a distributed computing system. The method attempts to minimize the cost incurred in replicating the resource at the processing sites and retrieving the resource by users of the system from the processing sites.
Several different approaches exist in the prior art for discovering information objects in Web caching schemes. Recent work has addressed the same resource allocation and discovery problems within the context of Internet services. Guyton and Schwartz (J. D. Guyton and M. F. Schwartz, “Locating Nearby Copies of Replicated Internet Servers,” Technical Report CU-CS-762-95, Department of Computer Science, University of Colorado-Boulder, February 1995; Proc. ACM SIGCOMM 95 Conference, Cambridge, Mass., August 1995, pp. 288-298) describe and analyze server location techniques for replicated Internet services, such as Network Time Protocol (NTP) servers and Web caches.
Guyton and Schwartz propose gathering location data with router support in two ways. In one method, routers advertise the existence or absence of replicated servers as part of their normal routing exchanges involving network topological information. Routers examine a distance metric for the advertised servers in a way that each router retains knowledge of at least the nearest servers. In this way, each router in an internetwork has enough knowledge to direct client requests to the nearest servers, without necessarily having to maintain knowledge of all the servers in the internetwork. In another method, servers poll routers for the content of their routing tables.
Guyton and Schwartz also describe a method for gathering location data using router probes without router support by means of measurement servers. According to this method, measurement servers explore the routes to the replicated serves providing services and content to the clients. When a client asks a measurement server for a list of nearby servers from which to request a service, the measurement server takes into account the route back to the client in deciding upon the list of servers that appear closer to the client.
One approach to object discovery consists in organizing Web caches hierarchically. In a hierarchical Web cache architecture, a parent-child relationship is established among caches; each cache in the hierarchy is shared by a group of clients or a set of children caches. A request for an information object from a client is processed at a lowest-level cache, which either has a copy of the requested object, or asks each of its siblings in the hierarchy for the object and forwards the request to its parent cache if no sibling has a copy of the object. The process continues up the hierarchy, until a copy of the object is located at a cache or the root of the hierarchy is reached, which consists of the servers with the original copy of the object.
One of the earliest examples of hierarchical Web caching was the Discover system (A. Duda and M. A. Sheldon, “Content Routing in Networks of WAIS Servers,” Proc. IEEE 14th International Conference on Distributed Computing Systems, June 1994; M. A. Sheldon, A. Duda, R. Weiss, J. W. O′Toole, Jr., and D. K. Gifford, “A Content Routing System for Distributed Information Servers,” Proc. Fourth International Conference on Extending Database Technology, March 1994), which provides associative access to servers; the user guides the refinement of requests.
Harvest (A. Chankhunthod, P. Danzing, C. Neerdaels, M. Schwartz, and K. Worrell, “A Hierarchical Internet Object Cache,” Proc. USENIX Technical Conference 96, San Diego, Calif., January 1996) and Squid (D. Wessels, “Squid Internet Object Cache,” http://www.squid.org, August 1998) are two of the best known hierarchical Web cache architectures. Harvest and Squid configure Web caches into a static hierarchical structure in which a Web cache has a static set of siblings and a parent. The Internet Caching Protocol or ICP (D. Wessels and K. Claffy, “Internet Cache Protocol (ICP), Version 2,” RFC 2186, September 1997) is used among Web caches to request information objects.
In the Harvest hierarchies, siblings and parents are configured manually in Web caches or proxies; this is very limiting and error prone, because reconfiguration must occur when a cache enters or leaves the system. A more general limitation of hierarchical Web caching based on static hierarchies is that the delays incurred in routing requests for information objects can become excessive in a large-scale system, and the latency of retrieving the information object from the cache with a copy of the object can be long, because there is no correlation between the routing of the request to a given cache in the hierarchy and the network delay from that cache to the requesting client. Furthermore, some Web caches may be overloaded with requests while others may be underutilized, even if they store the same objects.
In the WebWave protocol (A. Heddaya and S. Mirdad, “WebWave: Globally Load Balanced Fully Distributed Caching of Hot Published Documents,” Technical Report BU-CS-96-024, Boston University, Computer Science Department, October 1996; A. Heddaya and S. Mirdad, “WebWave: Globally Load Balanced Fully Distributed Caching of Hot Published Documents,” Proc. IEEE 17th International Conference on Distributed Computing Systems, Baltimore, Md., May 1997) Web caches are organized as a tree rooted at the server that provides the original copy of one object or a family of information objects; the leaves of the tree are the clients requesting the information objects, and the rest of the nodes in the tree are Web caches. The objective of the protocol is to achieve load balancing among Web caches; each Web cache in such a tree maintains a measurement of the load at its parent and children in the tree, and services or forwards the request to its parent automatically based on the load information. This approach reduces the possibility of overloading Web caches as in the Harvest approach to hierarchical Web caching; however, delays are still incurred in the propagation of requests from heavily loaded Web caches to their ancestors in the Web hierarchy.
Hash routing protocols (K. W. Ross, “Hash Routing for Collections of Shared Web Caches,” IEEE Network, Vol. 11, No. 6, November 1997, pp 37-44) constitute another approach to support object discovery in shared caches. Hash routing protocols are based on a deterministic hashing approach for mapping an information object to a unique cache (D. G. Thaler and C. V. Ravishankar, “Using Name-Based Mappings To Increase Hit,” IEEE/ACM Trans. Networking, 1998; V. Valloppillil and J. Cohen, “Hierarchical HTTP Routing Protocol,” Internet Draft, http://www.nlanr.net/Cache/ICP/draft-vinod-icp-traffic-dist-00.txt) to distribute the information objects (universal resource locator or URL in the case of the Web) among a number of caches; the end result is the creation of a single logical cache distributed over many physical caches. An important characteristics of this scheme is that information objects are not replicated among the cache sites. The hash function can be stored at the clients or the cache sites. The hash space is partitioned among the N cache sites. when a client requires access to an information object o, the value of the hash function for o, h(o), is calculated at the client or at a cache site (in the latter case the cache would be configured at the client, for example). The value of h(o) is the address of the cache site to contact in order to access the information object o.
The Cache Resolver is another recent approach to hierarchical Web caching (D. Karger, E. Lehman, T. Leighton, M. Levine, D. Lewin, and R. Panigrahy, “Consistent Hashing and Random Trees: Distributed Caching Protocols for Relieving Hot Spots on the World Wide Web,” Proc. 29th ACM Symposium on Theory of Computing (STOC 97), El Paso, Tex., 1997; D. Karger, Sherman, A. Berkheimer, B. Bogstad, R. Dhanidina, K. Iwamoto, B. Kim, L. Matkins, and Y. Yerushalmi, “Web Caching with Consistent Hashing,” Proc. 8th International World Wide Web Conference, Toronto, Canada, May 1999). This approach combines hierarchical Web caching with hashing and consists of two main tools, random cache trees and consistent hashing. A tree of Web caches is defined for each information object. When a browser (client) requires an information object, it picks a leaf of the tree and submits a request containing its identifier, the identifier of the object, the sequence of caches through which the request is to be routed if needed. A Web cache receiving a request determines if it has a local copy of the page and responds to the request if it does; otherwise, it forwards the request to the next Web cache in the path included in the request.
A Web cache starts maintaining a local copy of an information object when the number of requests it receives for the object reaches a predefined number. A client selects a Web cache by means of consistent hashing, which disseminates requests to leaves of the Web caching hierarchy evenly but, unlike traditional hashing techniques, need not redistribute an updated hash table every time a change occurs in the caching hierarchy (e.g., a new Web cache joins or a Web cache fails). Because caching is difficult to implement or add to existing Web browsers, the Cache Resolver approach implements the hashing in DNS servers modified to fit this purpose.
The remaining limitations with this approach stem from the continuing use of a hierarchy of Web caches and the need to implement a hashing function in either Web clients or DNS servers. Routing a request through multiple Web caches can incur substantial delays for clients to retrieve information objects that are not popular among other clients assigned to the same Web cache by the hashing function. Additional delays, even if small, are incurred at the DNS server that has to provide the address of the Web cache that the client should access. Furthermore, the DNS servers supporting the consistent hashing function must receive information about the loading of all the Web caches in the entire system, or at least a region of the system, in order to make accurate load-balancing decisions.
This DNS-based approach, without the use of hierarchies of Web caches, is advocated in the Akamai CDN solution (F. T. Leighton and D. M. Lewin, “Global Hosting System,” U.S. Pat. No. 6,108,703, Aug. 22, 2000). The “global hosting system” advocated by Akamai assumes that a content provider services an HTML document in which special URLs specifying a domain name specific to Akamai. When the client needs to obtain the IP address of the Web cache hosting the content specified in the special URL, the client first contacts its local DNS. The local DNS is pointed to a “top-level” DNS server that points the local DNS to a regional DNS server that appears close to the local DNS. The regional DNS server uses a hashing function to resolve the domain name in the special URL into the address of a Web cache (hosting server) in its region, which is referred to as the target Web cache in the present application, in a way that the load among Web caches in the region is balanced. The local DNS passes the address of that Web cache to the client, which in turn sends its request for the information object to that Web cache. If the object resides in the target Web cache, the cache sends the object to the client; otherwise, the object is retrieved from the original content site.
The global hosting system advocated by Akamai was intended to address problems associated with traditional load-balanced mirroring solutions in which a load balancer or a hierarchy of load balancers redirect requests to one of a few hosting sites to balance the load among such sites. Companies such as Cisco Systems of Santa Clara, Calif., F5 Networks, Inc. of Seattle, Wash., Resonate, Inc. of Sunnyvale, Calif., Nortel Networks of Brampton, Ontario, and Foundry Networks, Inc. of San Jose, Calif. currently provide examples of load-balanced solutions. The limitations of the global hosting system are inherent to the fact that the approach is, in essence, a DNS-based load-balanced mirroring solution. The global hosting system selects a target Web cache based entirely on the region that appears to favor the local DNS, which need not favor the client itself, and balances the load among Web caches without taking into account the latency between the Web caches and the clients. In the case of a cache miss, the information object has to be retrieved from the original content site, which means that latencies in the delivery of content can vary widely, unless the content is mirrored in all the caches of all regions.
Another alternative approach to hierarchical web caching and hash routing protocols consists of forwarding client requests for URLs using routing tables that are very similar to the routing tables used today for the routing of IP packets in the Internet (L. Zhang, S. Michel, S. Floyd, and V. Jacobson, “Adaptive Web Caching: Towards a New Global Caching Architecture,” Proc. Third International WWW Caching Workshop, Manchester, England, June 1998, B. S. Michel, K. Nikoloudakis, P. Reiher, and L. Zhang, “URL Forwarding and Compression in Adaptive Web Caching,” Proc. IEEE Infocom 2000, Tel Aviv, Israel, April 2000). According to this approach, which is referred to as “URL request forwarding” herein, Web caches maintain a “URL request routing table” and use it to decide how to forward URL requests to another Web caches when requested information objects are not found locally. The keys of the URL request routing tables are URL prefixes, which are associated with one or more identifiers to the next-hop Web caches or cache groups, and a metric reflecting the average delay to retrieve a request from a matching URL.
In this approach, an entry in the URL request routing table specifies a URL prefix and the next-hop Web cache towards an area or neighborhood of Web caches where the object resides. Ideally, a Web cache needs to know where a copy of a given object resides; however, because of the large number of objects (identified by URLs) that can be requested in a system, the URL request forwarding approach requires Web caches to be organized into areas or neighborhoods. All Web caches within the same area know the objects available in every other Web cache in the same area. In addition, for those objects that are not found in the area of a Web cache, the Web cache also maintains the next-hop Web cache towards the area in which a Web cache with the content resides.
Unfortunately, this approach has several scaling and performance limitations. First, requiring each Web cache to know all the Web caches where each object in the area resides incurs a large overhead, which is akin to the overhead of a traditional topology-broadcast protocol for IP routing, with the added disadvantage that the number of objects that can reside in an area can be much larger than the number of IP address ranges maintained in backbone routers of the Internet. Second, because Web caches only know about the next hop towards a URL that does not reside in a region, a request for an object that lies outside the area of a Web cache may traverse multiple Web-cache hops before reaching a Web cache in the area where an object is stored. This introduces additional latencies akin to those incurred in the caching hierarchies proposed in other schemes discussed above. Third, it is difficult to modify Web caches in practice to implement the mechanisms needed for the forwarding of URL requests.
To reduce the delays incurred in hierarchical Web caches, Tewari, Dahlin, Vin and Kay (R. Tewari, “Architectures and Algorithms for Scalable Wide-area Information Systems,” Ph.D. Dissertation, Chapter 5, Computer Science Department, University of Texas at Austin, August 1998; R. Tewari, M. Dahlin, H. M. Vin, and J. S. Kay, “Design Considerations for Distributed Caching on the Internet,” Proc. IEEE 19th International Conference on Distributed Computing Systems, May 1999) introduce hint caches within the context of a hierarchical Web caching architecture. According to this scheme, a Web cache maintains or has access to a local hint cache that maintains a mapping of an object to the identifier of another Web cache that has a copy of the object and is closest to the local hint cache. Web caches at the first level of the hierarchy maintain copies of information objects, while Web caches at higher levels only maintain hints to the objects. Hints are propagated along the hierarchy topology from the Web caches lower in the hierarchy to Web caches higher in the hierarchy. Furthermore, a Web cache with a copy of an object does not propagate a hint for the object. The limitation with this approach is that a Web caching hierarchy must still be established, which needs to be done manually in the absence of an automated method to establish the hierarchy, and the Web caching hierarchy must match the locality of reference by clients to reduce control overhead.
A number of proposals exist to expedite the dissemination of information objects using what is called “push distribution” and exemplified by Backweb, Marimba and Pointcast (“BackWeb: http://www.backweb.com/”’; ‘“Marimba: http://www.marimba.com/’”; “Pointcast: http://www.pointcast.com/’”). According to this approach, a Web server pushes the most recent version of a document or information object to a group of subscribers. The popular Internet browsers, Netscape Navigator and Internet Explorer™, use a unicast approach in which the client receives the requested object directly from the originating source or a cache. As the number of subscribers of a document or information object increases, the unicast approach becomes inefficient because of processing overhead at servers and proxies and traffic overhead in the network. The obvious approach to make push distribution scale with the number of subscribers consists of using multicast technology. According to this approach (P. Rodriguez and E. W. Biersack, “Continuous Multicast Push of Web Documents over The Internet,” IEEE Network Magazine, Vol. 12, No. 2, pp. 18-31, 1998), a document is multicasted continuously and reliably within a multicast group. A multicast group is defined for a given Web document and subscribers join the multicast group of the Web document they need to start receiving the updates to the document. A multicast group consists of the set of group members that should receive information sent to the group by one or multiple sources of the multicast group. The main shortcoming of this particular approach to push distribution are:                The portion of the Internet where subscribers are located must support multicast routing distribution.        A multicast address and group must be used for each Web document that is to be pushed to subscribers, which becomes difficult to manage as the number of documents to be pushed increases.        
Furthermore, Rodriguez, Biersack, and Ross (P. Rodriguez, E. W. Biersack, and K. W. Ross, “Improving the WWW: Caching or Multicast?,” Institut EURECOM 2229, Route Computer Networks and ISDN Systems, pp. 1-17 (Mar. 30, 1998) have shown that multicasting Web documents is an attractive alternative to hierarchical Web caching only when the documents to be pushed are very popular, caching distribution incurs less latency.
Kenner and Karush (B. Kenner and A. Karush, “System and Method for Optimized Storage and retrieval of Data on a Distributed Computer Network,” U.S. Pat. No. 6,003,030, Dec. 14, 1999) propose a method for expediting the delivery of information objects to end users. In this method, the end user site is equipped with special software in addition to the Web browser. This software consists of a configuration utility and a client program. The configuration utility is used to download a delivery site file specifying a list of the delivery sites (Web caches or originating Web servers) from which the information objects can be retrieved and a suite of tests that can be run to determine which delivery site to contact. The limitations with this approach stem from the fact that it is not transparent to end user sites. In particular, the end user site needs to run additional software; performance tests must be conducted from the end-user site to one or more delivery sites to decide which site to use; and when changes occur to the delivery sites, a new version of the delivery site file must be retrieved by the end-user site, or new performance tests must be conducted.
Another approach to helping select servers in a computer network (Z. Fei, S. Bhattacharjee, E. W. Zegura, and M. H. Ammar, “A Novel Server Selection Technique for Improving The Response Time of a Replicated Service” Proc. IEEE Infocom 98, March 1998, pp. 783-791) consists of broadcasting server loading information after a certain load threshold or time period is exceeded. The limitation of this approach is that, just as with topology-broadcast protocols used for routing in computer networks, the scheme incurs substantial overhead as the number of servers increases.
Still another approach to directing clients to hosting sites with requested information objects or services is the replica routing approach proposed by Sightpath, Inc. (D. K. Gifford, “Replica Routing,” U.S. Pat. No. 6,052,718, Apr. 18, 2000). According to the Replica Routing approach, an information object or service is replicated in a number of replica servers. The replica routing system redirects a client requesting the information object or service to a “nearby” replica of the object or service. In one approach, all replica routers know the replica advertisements from each of the replica servers in the system, which summarize information about their location and observations about the local internetwork topology and performance. Using this flooding of advertisements, a replica router discerns which replica server appears nearby any one client. However, requiring each replica router to receive the advertisements from every other replica server becomes impractical as the number of replica servers and replica routers increases.
To remedy this problem, replica routers are organized into a hierarchy, and replica advertisements are propagated only part way up such router hierarchy. A client request is routed to the root of the hierarchy and from there is forwarded down the hierarchy, until it reaches a replica router with enough knowledge about the replica's internetwork location to make an informed redirection decision. This approach has similar performance and scaling limitations as the prior approaches summarized above based on hierarchies of Web caches, flooding of information among caches or servers, and forwarding of requests over multiple hops.
Another recent approach to directing clients to hosting sites with requested information objects or services is the enhanced network services method by Phillips, Li, and Katz (S. G. Phillips, A. J. Li, and D. M. Katz, “Enhanced Network Services Using a Subnetwork of Communicating Processors,” U.S. Pat. No. 6,182,224, Jan. 30, 2001.). Insofar as directing clients to servers, the enhanced network services method is very similar to the gathering of location data with router support advocated by Guyton and Schwartz described previously. This technique is used by Cisco Systems of Santa Clara, Calif.
As in the Guyton and Schwartz approach, routers using the enhanced network services approach gather network topological data and also include as part of their normal routing exchanges information about the hosts that can provide content and services to clients; routers can then rank the hosts according to their relative distance in the network. In addition to data regarding hosts that can provide services, routers in the enhanced network services approach can include in their normal routing exchanges host information regarding logged-in users and willingness to pay for performing a designated service. This approach has a number of performance and scaling limitations just like other prior approaches summarized above. These limitations include the following.
The client in Cisco's enhanced network services method is pre-configured to contact a specific name server or router. In the case of a name server, the name server contacts the router on behalf of the client. In either case, the router determines the best server from the given list of server addresses. The router uses its routing tables (and server loads when available) to locate the best server. The assumption is that the proxy name server or router is close to the client in network terms. This assumption and preconfiguration are too restrictive. A more general solution is needed to enable an entity capable to provide enhanced network services to be contacted by any client to obtain best server information.
In contrast to the proposal by Guyton and Schwartz, the enhanced network services approach does not attempt to limit the amount of network topological information that routers need to exchange in order to direct clients to best qualified servers. Hence, the use of flooding of information among nodes providing enhanced network services and the forwarding of requests over multiple hops limits the scalability of the Cisco approach.
In Cisco's enhanced network services method, the client obtains the list of server addresses and provides it to the router for ordering them based on distances. The network latencies from the client to the servers are obtained from “traditional” routing tables. The server load information is then taken into account on demand at the router that receives the request. This imposes unnecessary functionality in the clients, because nodes in charge of enhanced services (not necessarily IP routers) could and should be more proactive in determining the best matches for clients in a completely transparent way from a client's standpoint. Furthermore, in Cisco's enhanced network services method, a router has route information only for itself to the servers. This is only partial information and, hence, the router is unable to determine distance from arbitrary clients to servers. The routers do not have the capability to combine network topology views from different routers. Constructing Internet topologies using inter-domain routing information (e.g., as provide from the BGP) is not performed by the routers involved in this process.